Imagine a vendor proposing an AI service that optimises an industrial process. The demonstration is convincing. Before the conversation reaches price, the operations leader asks: “What can we still run, understand and improve when this contract ends?”
That is a useful question even when buying the service is the right decision. It connects the promised improvement to a plant's long operating life, the organisation's accumulated knowledge and the responsibilities that remain with its people.
The intelligent industrial enterprise is often described through a future factory. I find it more useful to begin with that contract conversation. Intelligence becomes strategically valuable when the company can apply it to a recurring operating problem, judge its contribution and retain enough control to keep delivering.
Use the following framework to connect an industrial AI proposal to the decisions it requires. The 2036 paths are planning scenarios rather than adoption forecasts.
The process comes before the intelligence layer
An industrial business may have many promising tasks and very few well-defined process problems. A maintenance summary, production forecast and technical document assistant can each be useful. Their value depends on what currently limits availability, throughput, quality or service.
Start with the operating consequence. If downtime is the concern, identify which failures drive it and which decisions could prevent or shorten them. If quotation is slow, distinguish missing customer information from engineering capacity and approval delays. If quality losses are high, separate a detection problem from a process-control problem.
That work may point towards sensors, integration, rules, better maintenance routines or different decision rights before it points towards a language model. An AI component belongs in the design when its expected mechanism is clear. Otherwise, the programme accumulates demonstrations whose benefits cannot be connected to the result management cares about.
The AI Opportunity Map is intended to make that comparison possible. It treats process outcomes as the measure of success and allows a simpler intervention to win.
What the company needs to control
Consider a hypothetical maintenance-planning service for a group of plants. The supplier combines equipment history, technician notes and production schedules to recommend work.
The first question is whether the inputs are usable and authorised. Different plants may record the same failure differently. A missing history can be mistaken for a reliable asset. Technician notes may contain local abbreviations that the supplier's model does not understand. The company needs a way to expose those limits before trusting the recommendation.
The next question concerns action. Who approves a change to the maintenance schedule? Which recommendations are advisory, and which can change a system record? What happens when a recommendation conflicts with an equipment instruction or a production constraint? The accountable plant roles need explicit rights; an optimisation score does not settle that conflict.
Then examine continuity. Can the organisation retrieve its records and evaluation results in a usable form? What support is available when the service changes? Can it resume the necessary manual process? Who owns incident diagnosis across the plant, integration team and supplier?
Owning every model is one possible answer, and often an expensive one. Control can also come from clear contractual rights, documented interfaces, an independent evaluation set, trained operators and a practicable exit. The required degree of control should follow the process's importance and the consequences of disruption.
A capability stack with named owners
An industrial AI programme needs more than technical access. It needs a chain of responsibilities that reaches from the operating problem to the running service.
The process owner defines the result and accepts the changed work. Domain specialists define unacceptable errors and important exceptions. Data owners make the inputs traceable and usable. Technology teams support integration, availability and change management. Risk and security functions establish the controls appropriate to the deployment. Finance tests whether a claimed benefit has a credible route to realisation.
These responsibilities can be shared across plants. A common retrieval service or evaluation facility may reduce duplicated effort. Local teams still need a voice in the cases used to judge it. A standard that ignores a site's unusual operating conditions can make the system appear more consistent while making its recommendations less useful there.
The executive task is to resolve dependencies and trade-offs across these owners. Where necessary, assign a delivery leader with authority to coordinate them. Do not use that appointment to leave process accountability or incident response undefined.
Three ways 2036 could look
Scenario planning is useful when it changes today's investment. The following three paths are deliberately conditional.
In a baseline path, AI becomes a routine support layer for documentation, planning and analysis. Adoption is uneven because integration, data quality and local operating constraints remain difficult. Evidence for this path would include useful task improvements that translate slowly into wider process change. A sensible response would prioritise reusable data access and a small set of workflows with clear owners.
In a transformative path, several process-level improvements become reliable enough to coordinate across assets and sites. Planning, maintenance and quality decisions increasingly share current information and explicit action boundaries. Evidence would need to include sustained operating gains, effective exception handling and usable continuity arrangements. The response could involve deeper integration and changes to roles, but only where those conditions hold.
In a constrained path, economics, service reliability, workforce acceptance or operational limitations keep more systems advisory. Evidence might be persistent review burdens, weak performance on local exceptions or supplier costs that erase the benefit. The response would preserve useful assistance, reduce excessive commitments and fix enabling conditions where doing so has an independent value.
These paths can coexist inside one company. The same organisation may have mature document support, uncertain predictive maintenance and no justification for autonomous control of a consequential process. A single corporate maturity score conceals that difference.
Workforce design is part of the investment
Changing the task changes the job around it. A technician asked to evaluate a recommendation needs enough context and time to disagree. A planner responsible for exceptions needs a manageable queue and access to the source information. Removing routine work can also remove the experience through which people learn to recognise unusual cases.
Budget for those effects while designing the system. Training should include failure examples and the fallback, not just the interface. Assess whether the proposed review responsibility is realistic during a busy shift. Involve employees in identifying tasks that are burdensome and decisions that demand local judgment.
A workforce plan built around a forecast of how many jobs “AI will replace” is too coarse for this work. Map the actual tasks, changed responsibilities and demand for the released capacity. Savings, avoided hiring and improved service each require a different realisation plan.
The next investment should improve the options available later
Across the three scenarios, several investments remain useful: better equipment and process records, clear decision rights, accessible operating knowledge, representative evaluation and the ability to recover when a service fails. Those foundations also support improvements that use no AI.
Choose a process where the consequence matters and the evidence can be collected. Fund the next bounded intervention, agree what result would justify expansion and preserve a workable alternative. Before signing the contract, return to the opening question. The quality of the answer is a practical test of whether the proposal improves the industrial enterprise or merely adds another dependency.